springrollszxpAI makes data science so much easier and more accessible. It's what google/stackexchange was on steroids, and I think education as a whole needs to be rethunk. It's capable at carrying you through an entire undergraduate in any math heavy field (possibly also others, but I can't tell).
Missed this post somehow, but disagree here, at least for the statistics half/portion of data science. There's so much shit statistics advice out there that it's trained on, and I see this a lot as a TA, but if you don't already have a good understanding of the subject it's very easy for the AI to gaslight you into thinking something is true when it's not.
Here's an example:
https://i.imgur.com/kzVOZP4.png
This probably seems reasonable at first glance but this is just incorrect.
That being said models will always improve (wouldn't be surprised if the better AI's can get this right) but when you're first learning things it can be pretty misleading.
I agree that there is a lot of poor input data, however this is not really a counter-argument. Instead of getting a 10/10 score with AI you will get an 8/10 maybe. In fact, our faculty recently tested the performance of AI on a few of our exams, and the grade hovered around 9.0 (out of a 0-10 grading scale). At worst you can compare it to incorrect information you read in an article, stackexchange, lecture notes, or a book - if you don't understand the reasoning you would fail in all these cases as well.
The only topics it completely fails on in my experience are specialized topics which typically appear in your graduate courses and your PhD research.
wiitabix69420as a maths phd student i hover between "ai is useless for research level maths and anyone saying otherwise is buying into hype or clueless" and "i am terrified ai is going to replace my job in the future". i guess (hope) the real answer is it will end up as a useful assistant like it has become for programming but i dont think anyone is sure.
The real answer so far seems to lean toward the latter as far as I see. Springrolls makes a good argument above, that the end-user should still be capable of interpreting and understanding the outcome.
To me, this is no different from how data science has been in recent years. Writing bespoke code/analyses is quite easy with the resources available (even before LLMs), but interpreting these results correctly requires someone that is well-versed in statistics and the underlying (business) processes which generate the data. It only makes the end-user more efficient.